Efficient target control of complex networks based on ... - PLOS

RESEARCH ARTICLE

Efficient target control of complex networks based on preferential matching Xizhe Zhang1,2*, Huaizhen Wang2, Tianyang Lv3,4 1 Key Laboratory of Medical Image Computing of Northeastern University, Ministry of education, Shenyang, Liaoning, China, 2 School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China, 3 College of Computer Science and Technology, Harbin Engineering University, Harbin, Heilongjiang, China, 4 IT Center, National Audit Office, Beijing, China * [email protected]

a1111111111 a1111111111 a1111111111 a1111111111 a1111111111

OPEN ACCESS Citation: Zhang X, Wang H, Lv T (2017) Efficient target control of complex networks based on preferential matching. PLoS ONE 12(4): e0175375. https://doi.org/10.1371/journal.pone.0175375

Abstract Controlling a complex network towards a desired state is of great importance in many applications. Existing works present an approximate algorithm to find the input nodes used to control partial nodes of the network. However, the input nodes obtained by this algorithm depend on the node matching order and cannot achieve optimum results. Here we present a novel algorithm to find the input nodes for target control based on preferential matching. The algorithm elaborately arranges the matching order of the nodes to reduce the size of the input node set. The results on both synthetic and real networks indicate that the proposed algorithm outperforms the previous algorithm.

Editor: Yongtang Shi, Nankai University, CHINA Received: July 13, 2016 Accepted: March 6, 2017 Published: April 6, 2017

Introduction

Copyright: © 2017 Zhang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

The control of complex networked systems plays an important role in many nature and technology applications. According to control theory [1–3], a system is controllable if the system can be driven from any initial state to any desired state in finite time. The external control signals can be inputted into the system through some suitable selected nodes. The nodes which received independent external signals are called input nodes [4], controls [5] or driver nodes[6]. An input node is the first node of a control path which transmits the control signals. The input nodes, used to fully control the network, can be obtained by maximum matching of the network [7]. The unmatched nodes are the minimum set of input nodes (in short, MIS). Based on this framework, the researchers have analyzed the structural properties of MIS [8– 10], roles of nodes in control [5], and robustness of controllability [11]. The size of MIS is found to be tied to the degree distribution [6], and mainly dominated by the number of the source and sink nodes [5]. Furthermore, the possible input nodes which participate in at least one MIS are connected by the input adjacency [4], and they exhibit a surprising bifurcation phenomenon of the dense networks [12], which is rooted by the emergence of giant control component [4]. In many real control scenarios, only a small fraction of nodes need to be controlled. This is called target control [13]. To control the target community of a network, Piao et.al [14]

Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Funding: This work was supported by the Fundamental Research Funds for the Central Universities of China under grand number N140404011, the Natural Science Foundation of China under grant numbers 60903009, 71272216, 91546110, and the Special Program for Applied Research on Super Computation of the NSFCGuangdong Joint Fund (the second phase). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

1 / 10

Efficient target control of complex networks based on preferential matching

Competing interests: The authors have declared that no competing interests exist.

presented a method which used immune nodes to facilitate the control of target communities. To find the input nodes to control any specific target nodes, a recent work [13] presented an analysis framework to investigate the target control of complex networks. They proposed an approximate greedy algorithm (GA) based on multiple maximum matchings to obtain the input nodes used to control the target nodes. However, the GA can only find the approximate minimum set of input nodes. If there exists more than one maximum matching in the network, the results of the GA strongly depend on which maximum matching is selected. For example, the number of input nodes may vary over a large range [13] (Fig 1). Therefore, finding the minimum number of input nodes for target control is still an unsolved problem. Here, we present a novel algorithm for finding input nodes to control the target nodes of a network. In contrast to the previous approach, we elaborately arranged the matching order of the nodes and tried to reduce the total number of input nodes. The results on both synthetic and real networks showed that we obtained fewer input nodes than the previous approach.

Method Consider a linear time-invariant system, its states can be described by the following: 8 < dx ¼ Ax þ Bu dt : y ¼ Cx

ð1Þ

Where x(t) = (x1(t),. . ., xN(t))T represents the system’s state; u(t) = (u1(t),. . .,uM(t))T represents the input vector and y represents the output vector; A is the transpose of the adjacency matrix, B is the input matrix and C is the output matrix which defines the target nodes we want to control. Let the network representation of above system be G(V, E), where V is the nodes set and E

Fig 1. Illustration of random matching using GA. (A) A sample network G with target nodes {3, 6, 7, 8}; (B-D). Three MISs obtained by GA and their matching process, in which D1 = {1, 2, 5}, D2 = {1, 5}, D3 = {1}. The input nodes are obtained by the following process: 1. Construct a bipartite graph B (sub-bipartite graph 1) in which the right side contains all target nodes and the left side contains the nodes pointing to the target nodes; 2. Find a maximum matching of B and denote the matched nodes by M; and 3. Let M be the set of new target nodes, and repeat steps 1 and 2 until no new matched nodes are found. The differences between the three matching processes are highlighted by the blue shadow. https://doi.org/10.1371/journal.pone.0175375.g001

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

2 / 10

Efficient target control of complex networks based on preferential matching

is the edges set. For a target node set T, we say the system is target controllable if the states of the target node set T can be driven from any initial state to a desired final state [13]. In previous work [13], a k-walk theory was proposed, and this theory proved that in a treelike network, if a node has paths of different lengths to each target node, the node can control these target nodes. However, a single node cannot control all target nodes in many networks. Therefore, for more general networks with loops, previous work [13] proposed an approximate algorithm based on multiple maximum matchings to obtain the input nodes. The algorithm constructs a series of bipartite graphs B = {B1(T1,F1,E1),. . .,Bi(Ti,Fi,Ei)} by following procedures: 1. Let the set of target nodes be T1, find the set of in-neighbor nodes of T1 and denoted it as F1, construct bipartite graphs B1(T1,F1,E1), where E1 is the set of edges between nodes set T1 and F1;2. Let the F1 be the new target set T2 and repeat step 1 to get bipartite graph B2(T2,F2,E2); 3. Repeat above steps until the set of in-neighbor nodes of the current target set Ti is empty. After constructing the bipartite graphs, the algorithm finds the maximum matchings of each bipartite graph, and the union of the unmatched nodes of all bipartite graphs is the set of input nodes used to control the target nodes. The key idea of this algorithm is to find the maximum matching for each sub-bipartite graph. However, in most networks, the maximum matchings are not unique. Therefore, even for a simple network, the algorithm produced different results with different maximum matchings. For example, for the network shown in Fig 1A, the algorithm obtained three different input node sets: D1 = {1, 2, 5}, D2 = {1, 5} and D3 = {1}. The reason for the multiple results is that the maximum matchings used in the algorithm are different. For example, if we matched edge e(1!4) rather than e(2!4) in sub-bipartite graph 3, node 2 would not act as an input node, resulting in the input node set D2 = {1, 5}. If we match edge e(6!7) rather than e(5!7) in sub-bipartite graph 2, we obtain only one input node D1 = {1} to control the entire target node set. Therefore, to reduce the total number of input nodes, we need to select the appropriate maximum matching for each sub-bipartite graph. However, the number of unmatched nodes of each sub-bipartite graph is fixed because the maximum matchings of each bipartite graph have the same size. The only way to decrease the number of input nodes is to allow the input nodes of different sub-bipartite graphs to overlap with one another. For example, in Fig 1D, the unmatched node of all four sub-bipartite graphs is node 1, which decreases the total number of input nodes from three to one. To obtain the expected input nodes of each bipartite graph, we use the preferential matching [15] to find maximum matching of each bipartite graph. The preferential matching method arranges the matching order of the nodes based on a predefined queue, and ensure that the nodes in the rear of the queue have a high probability of being input nodes. The preferential matching method first constructs a series of sub-graphs based on the node queue, and then finds the maximum matching of each sub-graph until the maximum matching of the whole network is obtained. This iterative matching process ensures that the nodes in the front of the queue have a high probability to be matched. Therefore, the resulted input nodes are most likely the nodes in the rear of the queue. Therefore, the problem is selecting the appropriate input nodes of each sub-bipartite graph to reduce the total number of input nodes. Here we present the following strategies: 1. The input nodes of the current sub-bipartite graph should be overlapped with the input nodes of the previous sub-bipartite graph. This process will decrease the total number of input nodes. 2. The nodes that frequently appear in the matching graph (for all sub-bipartite graphs) should be input nodes with high priority, which will give the nodes in subsequent subbipartite graphs high probability to overlap with existing input nodes.

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

3 / 10

Efficient target control of complex networks based on preferential matching

Fig 2. Illustration of preferential matching for target control. (A). A sample network G with target nodes {3, 6, 7, 8}. (B). Matching graph for target nodes {3, 6, 7, 8}. (C). Matching sequence of nodes based on their counts in the matching graph. The counts for node sequence {n1,n2,n6,n4,n5,n7,n3,n8} are {4,3,3,3,2,2,1,1}. https://doi.org/10.1371/journal.pone.0175375.g002

Fig 2 illustrates these strategies on an example network. For the network shown in Fig 2A, we construct a matching graph (MG) that starts from the target nodes and iteratively adds the parent nodes of current nodes to the graph, until no more nodes are added. We count the frequency with which each node appeared in the matching graph and arrange the nodes in ascending order of frequency. For example, Fig 2B shows the matching graph of Fig 2A, and the counts of nodes are n1 = 4, n2 = 3, n3 = 1, n4 = 3, n5 = 2, n6 = 3, n7 = 2 and n8 = 1, respectively. Therefore, the matching sequence of nodes should be {n8, n3, n7, n5, n4, n6, n2, n1} according their counts by ascending order. For each sub-bipartite graph of MG, we used this matching sequence to find input nodes. Overall, for a network G and target node set T, the algorithm based on preferential matching (PM) for finding input nodes consists of the following steps: 1. For target node set T, construct bipartite graph B1(F, T), where F are the node sets pointing to target node set T. 2. Let F be the new target node set. Repeat step 1 to construct bipartite graph Bi(F, T) until no more nodes are found. Define the matching graph M(T) = {B1, B2, . . ., Bi}. 3. For each node in M, compute their counts f(n), arrange the nodes by ascending order of f (n) and let Q be the matching sequence. 4. For a sub-bipartite graph of M, find the maximum matching based on preferential matching using node sequence Q. Let D = {d1,d2,. . .,di} be the set of input nodes. Rearrange the node sequence by putting the nodes of D in the rear of Q. 5. Repeat step 4. Find input nodes Di of sub-bipartite graph Bi. The final set of input nodes to control the target nodes is D = [ Di.

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

4 / 10

Efficient target control of complex networks based on preferential matching

Result To quantify the efficiency of the algorithm, we evaluated the fraction of input nodes nD = ND/ N based on a PM algorithm and GA [13]. We used the following two different schemes for target node selection: 1. Random selection scheme: Select nodes from the network uniformly at random as targets, until reaching the expected target fraction f. 2. Local selection scheme: Randomly select a seed node, and expand the node based on a breadth-first search (BFS) tree, until reaching the expected target fraction f. Fig 3 shows the results of scale-free networks [16] with N = 104. For different target node fractions f 2 [0,1], the PM algorithm always has better performance than the GA in both target node selection schemes. Furthermore, the difference in the values of nD obtained by PM and GA, |ΔnD| = |nD-GA- nD-PM|, increases with the fraction of target nodes f, suggesting that the PM algorithm is more efficient in controlling large fractions of target nodes.

Fig 3. Efficiency analysis of the target control algorithm for two synthetic networks. (A-B). For the scale-free networks with N = 104 and = 5.2, we show the density of input nodes as a function of the fraction of target nodes. The results are computed based on 100 network instances with the same average degree. (A) Results of the local selection scheme and (B) Results of the random selection scheme. (C-D) For the scale-free networks with N = 104 and = 13, we show the density of input nodes as a function of the fraction of target nodes. (C) Results of the local selection scheme and (D) Results of the random selection scheme. For each network, we compute the fraction of input nodes nD based on the preferential matching and the greedy algorithm. For the greedy algorithm, the nD is computed based on the results of 100 random experiments. https://doi.org/10.1371/journal.pone.0175375.g003

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

5 / 10

Efficient target control of complex networks based on preferential matching

Fig 4. The efficiency of the algorithm for different average degree . (A-B). For a scale-free network with N = 104 and target node fraction f = 0.3, we show (A) the density of input nodes versus , based on the local selection scheme, and (B) the density of input nodes versus , based on the random selection scheme. (C-D). For an ER random network with N = 104 and target node fraction f = 0.3, we show (C) the density of input nodes versus , based on the local selection scheme, and (D) the density of input nodes versus , based on the random selection scheme. For each average degree , the fraction of input nodes nD is computed based on the average results of 100 networks. https://doi.org/10.1371/journal.pone.0175375.g004

Next, we analyzed nD with different average degrees . Fig 4 shows the results for both the scale-free networks and ER random networks based on local and random target selection schemes. The PM algorithm obtains lower nD than GA in all networks. Note that the variations of nD for the local selection scheme of target nodes are much larger than those variations for the random selection scheme, suggesting that there are many input nodes set to control target nodes that are locally connected. We also evaluated the performance of the PM algorithm in real networks. The networks are selected based on diversity of topological structure and include food web, transcription, citation, and Internet networks. The results are shown in Table 1 and Fig 5. For all networks and fractions of target nodes in both random and local schemes, the PM algorithm outperforms the GA.

Discussion The controllability of complex networks is of great importance in many applications. Controlling a small fraction of target nodes is a common task in many real control scenarios. Here we proposed a novel algorithm based on preferential matching to reduce the number of input

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

6 / 10

Efficient target control of complex networks based on preferential matching

Table 1. Results for the real networks analyzed in the paper. Type

Name

Food Web

N

L



Random selection

Local selection

npd

nrd

npd

nrd

Mangrove[17]

97

1492

30.76

3.09%

5.05%

4.12%

6.27%

Silwood[18]

154

370

4.81

30.52%

31.27%

24.03%

25.37%

Neuronal

C. elegans[19]

306

2345

15.33

7.07%

7.97%

4.38%

Transcription

E. coli[20]

423

578

2.73

39.95%

43.04%

32.39%

38.02

TRN-Yeast-1[21]

4441

12873

5.80

45.49%

45.68%

43.75%

44.13%

TRN-Yeast-2[22]

688

1079

3.14

39.68%

40.67%

31.40%

33.33%

Prison inmate[23, 24]

67

182

5.43

13.43%

17.51%

11.94%

13.84%

WikiVote[25]

7115

103689

29.15

39.68%

40.33%

37.26%

38.37%

s208a[22]

122

189

3.10

9.02%

14.75%

6.56%

10.03%

s420a[22]

252

399

3.17

8.33%

14.43%

3.57%

7.33%

Trust Electronic circuits

5.55%

s838a[22]

512

819

3.20

6.64%

12.10%

1.95%

5.77%

Citation

ArXiv-HepTh[26]

27770

352807

25.41

13.41%

15.10%

7.62%

10.03%

Kohonen[27]

4470

12731

5.70

17.90%

21.11%

9.49%

13.26%

WWW

Political blogs[28]

1224

16718

27.32

10.87%

13.26%

7.84%

10.99%

Internet

p2p-1[29]

10876

39994

7.35

3.51%

9.41%

0.77%

4.80%

p2p-2[29]

8846

31839

7.20

4.70%

10.44%

2.32%

6.11%

p2p-3[29]

8717

31525

7.23

4.73%

10.52%

2.44%

7.48%

UClonline[30]

1899

20296

21.38

8.48%

11.14%

3.95%

6.91%

Social network

Facebook_0[31]

347

5038

29.04

1.80%

2.69%

0.30%

0.60%

Facebook_107[31]

1912

53498

55.96

0.10%

0.19%

0.10%

0.10%

Facebook_348[31]

572

6384

22.32

0.89%

1.79%

0.45%

0.45%

For each network, we show its type, name, number of nodes (N) and edges (L), average degree , density of input nodes (npd) based on preferential matching, and density of input nodes (nrd) based on random matching. The fraction of target nodes f = 0.5. https://doi.org/10.1371/journal.pone.0175375.t001

Fig 5. Results for real networks. We show the fraction of input nodes and the fraction of target nodes. The PM method always achieves better performance in both the local and random target node selection schemes. https://doi.org/10.1371/journal.pone.0175375.g005 PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

7 / 10

Efficient target control of complex networks based on preferential matching

nodes. Our algorithm has the same main steps as the previous algorithm [13], based on multimaximum matching of the induced bipartite graphs. However, we elaborately arranged the matching order of the nodes, which can significantly reduce the number of resulting input nodes. However, our algorithm still cannot guarantee the optimum result. Future work should focus on finding an efficient and precise method to reduce the number of input nodes.

Supporting information S1 Fig. Efficiency analysis of the target control algorithm for three scale-free networks with N = 104. We show the density of input nodes as a function of the fraction of target nodes based on local and random schemes. For each network, we compute the density of input nodes nD based on the preferential matching and the greedy algorithm. For the greedy algorithm, the nD is computed based on the results of 100 random experiments. (TIF) S2 Fig. The density of input nodes nD versus the fraction of target nodes f of real networks. We show the results of ArXiv-HepTh, C.Elegans, Kohonen and Facebook_0 networks. (TIF) S3 Fig. The density of input nodes nD versus the fraction of target nodes f of real networks. We show the results of P2P-2, P2P-3, S208, S420 and S838 networks. (TIF)

Author Contributions Conceptualization: XZZ TYL. Data curation: HZW. Formal analysis: XZZ HZW. Funding acquisition: XZZ TYL. Investigation: XZZ HZW. Methodology: XZZ HZW. Project administration: XZZ. Software: HZW. Supervision: XZZ. Validation: XZZ HZW. Visualization: HZW XZZ. Writing – original draft: XZZ. Writing – review & editing: XZZ.

References 1.

Kalman RE. Mathematical Description of Linear Dynamical Systems. Journal of the Society for Industrial & Applied Mathematics. 1963; 1(2):152–192.

2.

Luenberger DG. Introduction to Dynamic Systems: Theory, Models, & Applications. Proceedings of the IEEE. 1979; 69(9):1173.

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

8 / 10

Efficient target control of complex networks based on preferential matching

3.

Slotine JJE, Li W. Applied nonlinear control. Beijing: China Machine Press; 2004.

4.

Zhang X, Lv T, Pu Y. Input graph: the hidden geometry in controlling complex networks. Scientific Reports. 2016; 6:38209. https://doi.org/10.1038/srep38209 PMID: 27901102

5.

Ruths J, Ruths D. Control profiles of complex networks. Science. 2014; 343(6177):1373–1376. https:// doi.org/10.1126/science.1242063 PMID: 24653036

6.

Liu YY, Slotine JJ, Barabasi AL. Controllability of complex networks. Nature. 2011; 473(7346):167–173. https://doi.org/10.1038/nature10011 PMID: 21562557

7.

Murota K. Matrices and Matroids for Systems Analysis. Berlin: Springer Science & Business Media; 2000.

8.

Jia T, Baraba´si AL. Control capacity and a random sampling method in exploring controllability of complex networks. Scientific Reports. 2013; 3:2354. https://doi.org/10.1038/srep02354 PMID: 23912679

9.

Jia T, Posfai M. Connecting core percolation and controllability of complex networks. Scientific Reports. 2014; 4:5379. https://doi.org/10.1038/srep05379 PMID: 24946797

10.

Po´sfai M, Liu YY, Slotine JJ, Baraba´si AL. Effect of correlations on network controllability. Scientific Reports. 2012; 3:1067.

11.

Pu CL, Pei WJ, Michaelson A. Robustness analysis of network controllability. Physica A Statistical Mechanics & Its Applications. 2012; 391(18):4420–4425.

12.

Jia T, Liu YY, Cso´ka E, Po´sfai M, Slotine JJ, Baraba´si AL. Emergence of bimodality in controlling complex networks. Nature Communications. 2013; 4:2002. https://doi.org/10.1038/ncomms3002 PMID: 23774965

13.

Gao J, Liu YY, D’souza RM, Baraba´si AL. Target control of complex networks. Nature Communications. 2014; 5:5415. https://doi.org/10.1038/ncomms6415 PMID: 25388503

14.

Piao X, Lv T, Zhang X, Ma H. Strategy for community control of complex networks. Physica A Statistical Mechanics & Its Applications. 2015; 421:98–108.

15.

Zhang X, Lv T, Yang XY, Zhang B. Structural controllability of complex networks based on preferential matching. PLoS One. 2014; 9(11):e112039. https://doi.org/10.1371/journal.pone.0112039 PMID: 25375628

16.

Baraba´si AL, Albert R. Emergence of scaling in random networks. Science. 1999; 286(5439):509–512. PMID: 10521342

17.

Ulanowicz RE, DeAngelis DL. Network analysis of trophic dynamics in south florida ecosystems. US Geological Survey Program on the South Florida Ecosystem. 2005;114.

18.

Montoya JM, Sole´ RV. Small world patterns in food webs. Journal of Theoretical Biology. 2002; 214 (3):405–412. https://doi.org/10.1006/jtbi.2001.2460 PMID: 11846598

19.

Watts DJ, Strogatz SH. Collective dynamics of ‘small-world’networks. Nature. 1998; 393(6684):440–2. https://doi.org/10.1038/30918 PMID: 9623998

20.

Shen-Orr S, Milo R, Mangan S, Alon U. Network motifs in the transcriptional regulation network of Escherichia coli. Nature Genetics. 2002; 31(1):64–68. https://doi.org/10.1038/ng881 PMID: 11967538

21.

Bu D, Zhao Y, Cai L, Xue H, Zhu X, Lu H, et al. Topological structure analysis of the protein–protein interaction network in budding yeast. Nucleic acids research. 2003; 31(9):2443–2450. PMID: 12711690

22.

Milo R, Shen-Orr S, Itzkovitz S, Kashtan N, Chklovskii DB, Alon U. Network motifs: Simple building blocks of complex networks. Science. 2002; 42(6821):285–298.

23.

Milo R, Itzkovitz S, Kashtan N, Levitt R, Shen-Orr S, Ayzenshtat I, et al. Superfamilies of evolved and designed networks. Science. 2004; 303(5663):1538–1542. https://doi.org/10.1126/science.1089167 PMID: 15001784

24.

Van Duijn MAJ, Zeggelink EPH, Huisman M, Stokman F, Wasseur FW. Evolution of sociology freshmen into a friendship network. Journal of Mathematical Sociology. 2003; 27(2–3):153–191.

25.

Leskovec J, Lang KJ, Dasgupta A, Mahoney MW. Community structure in large networks: Natural cluster sizes and the absence of large well-defined clusters. Internet Mathematics. 2009; 6(1):29–123.

26.

Leskovec J, Kleinberg J, Faloutsos C. Graphs over time: densification laws, shrinking diameters and possible explanations. Proceedings of the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining; 2005: 177–187.

27.

Handcock MS, Hunter D, Butts CT, Goodreau SM, Morris M. Statnet: An R package for the Statistical Modeling of Social Networks. 2003. http://www.csde.washington.edu/statnet.

28.

Adamic LA, Glance N. The political blogosphere and the 2004 US election: divided they blog. Proceedings of the 3rd international workshop on Link discovery; 2005: 36–43.

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

9 / 10

Efficient target control of complex networks based on preferential matching

29.

Leskovec J, Kleinberg J, Faloutsos C. Graph evolution: Densification and shrinking diameters. ACM Transactions on Knowledge Discovery from Data (TKDD). 2007; 1(1):2–41.

30.

Opsahl T, Panzarasa P. Clustering in weighted networks. Social Networks. 2009; 31(2):155–163.

31.

Mcauley JJ, Leskovec J. Learning to discover social circles in ego networks. Advances in Neural Information Processing Systems; 2012:539–547.

PLOS ONE | https://doi.org/10.1371/journal.pone.0175375 April 6, 2017

10 / 10